Pedestrian Crashes at High-Speed Intersections: Applying AI for Crash Narrative Analysis to Identify Common and Edge Cases
Project Description
High-speed intersections, with speed limits of 35 mph or higher, and long crossings are particularly risky for pedestrians. Tragically, in 2022, 983 pedestrians were killed at signalized intersections, representing about 16% of all pedestrian fatalities. Intersections are concerning due to pedestrians' difficulty navigating them and the numerous conflict points they present. The proposed research addresses a critical gap: understanding pedestrian crash descriptors using structured and unstructured (narrative) data on high-speed intersections and identifying simple and complex (or edge) cases—unusual or extreme crashes that deviate substantially from the typical ones. Complex cases represent exceptional circumstances with many contributing factors. Understanding them will help inform current pedestrian safety strategies and help improve how autonomous vehicle algorithms anticipate potential pedestrian conflicts. The research question is: What are the different types of pedestrian crashes and injuries at intersections, and what are their complexity levels?
To answer this question, the team will use data from crashes from Tennessee’s Integrated Traffic Analysis Network (TITAN) and Wisconsin’s WisTransPortal, for which the research team has access to police crash narratives. The team will separate pedestrian crashes at high-speed signalized intersections (with speed limits of 35 mph or higher) and compare them to those at intersections with speed limits of less than 35 mph. The team will use AI to develop high-quality, detailed crash descriptors from the narratives of police reports and quantitative crash data. Natural language processing and feature extraction techniques will categorize pedestrian crashes into specific types based on detailed pre-crash actions, human errors, and circumstances obtained from structured and unstructured data. The study will identify edge cases and relevant safety countermeasures (e.g., conflict reductions) while providing a nuanced understanding of crash circumstances (relative to current practice).
The study will create a unique and comprehensive crash database that can provide deep insights into the range of injuries, crash attributes (e.g., crash location within the intersection or pedestrian and driver actions), precrash positions, driver and pedestrian impairment, and roadway conditions, and design (e.g., visibility, number of lanes, pedestrian crossing facilities). The study will apply rigorous analysis methods, including unsupervised learning techniques to identify complex cases and inference-based frequentist methods to quantify key correlates of crash injuries. Cluster analysis, specifically through hierarchical or k-means techniques, will differentiate complex crash cases from more common ones, effectively isolating extreme cases deviating from typical patterns.
In addition to highlighting the issue of pedestrian crashes at intersections and their correlates, a unique aspect of this study is the identification of complex cases. By doing so, the study aims to uncover the underlying patterns and risk factors that contribute to complex and unusual pedestrian crashes at intersections. Rather than focusing solely on the common crash situations, considering a wide range of possibilities and using a unique database helps us understand and address common and rare cases for high-speed and low/medium-speed intersections (especially relevant, given the adoption of vehicle automation and higher safety standards), ensuring a safer environment for vulnerable road users.
Outputs
The results of this project will generate new methods, data resources, and analytical tools to improve pedestrian safety at high-speed intersections. Specifically, the study will produce a comprehensive crash database that integrates structured police-reported data with unstructured narrative data from TITAN and WisTransPortal, creating a richer and more detailed source than currently available. Using natural language processing and feature extraction, the project will develop a process for systematically categorizing pedestrian crashes by type, contributing factors, and complexity level. Advanced analytical methods, including unsupervised learning, will enable the identification of edge cases and anomaly detection, providing novel insights into unusual or extreme crashes that are often overlooked in standard analyses. The outputs will include methods, datasets, and code that researchers and practitioners can use. These resources, along with the Final Research Report, will equip planners, engineers, and AV developers with tools to anticipate rare crash contexts, refine safety countermeasures, and design safer intersections for vulnerable road users. Notably, the project efforts will build on initial findings by the research team, suggesting that more unusual contexts, such as combinations of poor lighting, high-speed roads, and rural settings, are associated with more severe pedestrian crashes. Additionally, crash types, such as“Crossing Expressway” and “Walking Along Roadway,” were especially more sensitive to these contexts, while routine types, such as bus-related crashes, were less affected.
Outcomes/Impacts
The final report from this research will provide detailed methods, findings, and safety recommendations, serving as a critical resource for traffic management professionals and infrastructure designers. By developing a comprehensive pedestrian crash database and advanced analytical methodologies, the project will transform the understanding and approach to high-speed intersection safety. These outputs will allow policymakers and engineers to implement more targeted safety measures and traffic management strategies. In addition, the introduction of novel algorithms and safety assessment tools will strengthen technological capabilities and support regulatory and legislative improvements, setting higher standards for AV algorithms and roadway design. Additionally, while the project emphasizes rare and complex crash cases, the framework also allows comparisons with more typical cases, helping decision-makers weigh where to focus resources for the greatest safety impact. Although initial research and development costs may be significant, the long-term benefits, such as reduced crashes involving vulnerable road users and lower healthcare expenses, will deliver substantial economic and social value. Ultimately, this research will enhance the safety, reliability, and cost-effectiveness of the transportation system, creating safer environments for all road users.
Dates
12/1/2025 to 11/30/2026
Universities
University of Tennessee Knoxville
University of Wisconsin Milwaukee
Principal Investigator
Asad Khattak
akhattak@utk.edu
https://orcid.org/0000-0002-0790-7794
Robert Schneider
rjschnei@uwm.edu
https://orcid.org/0000-0002-6225-3615
Project Partners
The University of Tennessee, Knoxville
Center for Transportation Research
University of Wisconsin-Milwaukee
Department of Urban Planning
Research Project Funding
Federal: $59,117
Non-Federal: $26,426
Contract Number
69A3552348336
Project Number
25UTK02
Research Priority
Promoting Safety